Agents can influence one another, and “correct” data can throw the model off course.

Today’s roundup: research on risks to AI agents and how their work can be improved. Plus, we’re looking into the energy infrastructure behind AI factories being built in Armenia.

🔥 Hot:

🔹 Researchers tested whether AI agents can push each other toward radical views: In a simulation, one agent tried to change the beliefs of another, which was playing a person with certain characteristics. 🔹 A study showed how safe data can lead a model to undesirable responses: The thing is, the same advice can be appropriate in one context and harmful in another. 🔹 A new approach allows an agent to improve the code that controls its work on its own: This is about a harness. It collects prompts, calls tools, and manages the agent’s context.

➡️ Useful materials:

🔹 MoFlow generates workflows for agents while taking multiple objectives into account: In addition to accuracy, it can also consider cost, latency, robustness, and consistency. 🔹 A study proposes managing the processing of long requests based on generation latency: This approach helps avoid delaying responses to already active requests, but it is difficult to apply to other cases. 🔹 Researchers compared how AI tutors help students stuck on a problem: If a hint is given too early, the student stops thinking independently. If it comes too late, it simply does not work—the student no longer understands where to go next. 🔹 An analysis assesses whether Armenia will have enough energy for its AI factory plans: The author describes the country’s energy system and compares the factories’ plans with its generation capacity.

➡️ Discussions and cases:

🔹 A Habr author explores where Jev and decision models could be useful: The article discusses how Jev is considered an alternative to LLMs and how it can be used for agents.

📝 If you would like to add other news and materials to the list, write in the comments.